truth-maintenance
Skill17 starsChanged 2 days agoApache-2.0View on GitHub
Coordinate a full truth-maintenance pass for stellar-raven-codemode across live drift, evals, golden questions/answers, and improvements/issues/PR follow-up. Use when asked to ensure repo truth is current, review the current state of upstream findings or PRs the repo opened, run a multi-lane maintenance round, or orchestrate CI-like review work across several agents instead of one monolithic context window.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- audit-reviewabilitySkill
- cloudflare-observability-reviewSkill
- golden-truthSkill
- improvements-pipelineSkill
- live-drift-resolutionSkill
- retrieval-system-auditSkill
- run-evalsSkill
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 5 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Use it when the work crosses lanes or the request is "is repo truth current?".
- Distinctive: line 27
only the lanes that question needs, and for each its commands or probes, expected artifact, and stop rule.
- Distinctive: line 32
Run one session when the lanes are small.
- Distinctive: line 75
Stop when each lane in scope has a verdict with evidence.
- Distinctive: line 81
Every recorded GitHub URL was read back.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Truth maintenance
This skill coordinates; the lane skills do the work:
| Lane | Skill | Owns |
| Drift | live-drift-resolution | Inventory, catalog, spec, and op-class drift |
| Eval | run-evals | Instrument choice, spend, and result review |
| Golden | golden-truth | Every golden question, answer, key fact, avoid, and note change |
| Improvements | improvements-pipeline | Findings, intake, probes, and upstream issue and PR state |
Use it when the work crosses lanes or the request is "is repo truth current?". For one drift issue or one stale case, use the lane skill directly.
1. Scope the round
Size the work before you start it. Write in the ledger
(.agents/rounds/<YYYY-MM-DD>-truth-maintenance.md):
- the question the round answers: stale-queue triage, drift, upstream follow-up, golden health, or release readiness;
- only the lanes that question needs, and for each its commands or probes, expected artifact, and stop rule;
- the expected time and cost. Lanes run here on
hcwith subscription CLI agents; any paid eval follows therun-evalsspend rule.
Run one session when the lanes are small. Spawn a lane agent (AGENTS.md "Coordination") only when
the lane is independent and large enough to pay for the handoff. Workers research and append to the
ledger under their own heading; the coordinator owns repo edits unless a write set is narrow,
explicit, and disjoint. A reviewer re-derives from files, diffs, live probes, or GitHub state, not
from the coordinator's summary.
Ledger headings: Scope, Lane plan, one verdict section per lane in scope, Own-repo todos,
Decisions, Final checklist.
2. The stale-gospel queue
This skill owns golden freshness as a standing due-date queue. Every scheduled case has a
truth.reverifyBy date, and npm run eval:qa:lint -- --stale fails on a past-due case in PR CI
and the daily refresh.
- When the gate fires, send each past-due case through
golden-truth: re-verify it (updatetruth.verified,truth.asOf, andreverifyBy), or record a dated extension with a root cause. The gospel-change lint does not require evidence when onlyreverifyBychanges, so a reviewer checks the evidence for either action. - Stagger new dates by quarter across cases. A wall of identical dates is a review finding.
- Source sentinels fire early. The daily
golden-sentinelsworkflow keeps one labeled issue open while an anchor drifts. Queue each drifted case ahead of itsreverifyBydate. A drift is a trigger, never evidence; when the fact holds and only the wording moved, re-author the anchor. - Look ahead. Fold cases due in the next two to four weeks into the current round.
- Corpus health, when the round asks for it: scan
eval/qa/consistency-register.jsonclusters for cross-question contradictions, sample a refute-then-repair sweep weighted toward freshness-sensitive, numeric, and non-confirmedcases, and re-verify anydateContingentTrapswhose trigger has passed. Fixes go throughgolden-truth.
3. Reconcile across lanes
- Drift changed a fact that a golden or eval case consumes.
- An eval failure implies a new or updated finding.
- An upstream issue or PR resolution needs a live re-run of the original trigger. A closed issue or merged PR is not proof.
- A golden change needs its root cause in
improvements/or.agents/TODO.md.
Track improvements, issues, and PRs in the state table from improvements-pipeline ("Follow up on
filed findings"). A fixed-upstream finding retires only after a second reviewer repeats the live
trigger. Declined findings stay active while the trigger reproduces.
4. Close
Stop when each lane in scope has a verdict with evidence. Then confirm:
- Each lane ran its own final gates; the ledger records commands, result stamps, URLs, and commits.
npm run eval:qa:lint -- --staleis green and newreverifyBydates are staggered.- Improvements index, lint, and probes ran if
improvements/changed; every retired finding has a completeresolved.jsonreceipt and no stale references. - Every recorded GitHub URL was read back; untouched open issues stayed silent.
- Open upstream refs have a next action or a dated
.agents/TODO.mdentry that names a concrete check. - Lanes you started are finished and their output is incorporated.
npm run secrets:scan -- --treeran before any commit with generated or copied upstream content.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

